Validation of a daily satellite-derived Antarctic sea ice velocity product: impacts on ice kinematics

Antarctic sea ice kinematic plays a crucial role in shaping the polar climate and ecosystems. Satellite passive 10 microwave-derived sea ice motion data have been used widely for studying sea ice motion and deformation processes, and provide daily, global coverage at a relatively low spatial-resolution (in the order of 60×60 km). In the Arctic, several validated data sets of satellite observations are available and used to study sea ice kinematics, but far fewer validation studies exist for the Antarctic. Here, we compare the widely-used passive microwave-derived Antarctic sea ice motion product by Kimura et al. (2013) with buoy-derived velocities, and interpret the effects of satellite observational configuration on the 15 representation of Antarctic sea ice kinematics. We identify two issues in the Kimura et al. (2013) product: (i) errors in two large triangular areas within the eastern Weddell Sea and western Amundsen Sea relating to an error in the input satellite data composite, and (ii) a more subtle error relating to invalid assumptions for the average sensing time of each pixel. Upon rectification of these, performance of the daily composite sea ice motion product is found to be a function of latitude, relating to the number of satellite swaths incorporated (more swaths further south as tracks converge), and the heterogeneity of the 20 underlying satellite signal (brightness temperature here). Daily sea ice motion vectors calculated using ascendingand descending-only satellite tracks (with a true ~24 h time-scale) are compared with the widely-used combined product (ascending and descending tracks combined together, with an inherent ~39 h time-scale). This comparison reveals that kinematic parameters derived from the shorter time-scale velocity datasets are higher in magnitude than the combined dataset, indicating a high degree of sensitivity to observation time-scale. We conclude that the new generation of “swath-to25 swath” (S2S) sea ice velocity datasets, encompassing a range of observational time scales, is necessary to advance future research into sea ice kinematics.

Recently sea ice motion has been estimated from PM-derived T B partial-overlap swath pairs from a wide range of temporal baselines (from one orbit of ~90 minutes, to 3 days). This approach is referred as "swath-to-swath" (S2S) (Lavergne et al., 2020). For S2S, the time base for derived motion is a function of the temporal separation of individual swath in each pair, 115 hence differs between pairs (Lavergne et al., 2020). As such, the raw S2S-derived motion swath do not constitute a regular, full coverage time series. However, since S2S most accurately provides ice motion from induvial swaths across a range of time bases, Lavergne et al. (2020) concluded that sea ice motion based on S2S is more accurate than that from DM.
We hypothesise that a greater number of satellite swaths used for averaged maps of satellite signals (e.g., TB DMs) may 120 negatively impact the performance of PM-derived sea ice motion vectors. Here, we investigate this hypothesis by performing comparisons against drifting buoy-derived motion in different latitude ranges with the expectation that DM-based KIMURA dataset performs better at lower latitudes, where swath overlap averaging is lower. We note that the OSI SAF sea ice motion product is derived by averaging over two days to increase the signal-to-noise ratio of this operational product (Lavergne, 2016). This extended averaging interval arises, in part, due to OSI SAF integrating multiple sensors. Because of the extended time base of the OSI SAF product, direct comparison with the NSIDC and KIMURA datasets in a kinematic parameters context is not possible. Hence the OSI SAF dataset is not considered further in this study, although implications of a longer temporal baseline are considered in the discussion. Version 3 of NSIDC dataset is known to display a high degree of autocorrelation in neighbouring vectors (Szanyi et al., 2016) which would be detrimental to the derivation of kinematic parameters. Version 4 of the NSIDC product has reduced this autocorrelation to some extent (Tschudi et al., 2019), however 130 this product still assimilates from several instruments, so retrieval of kinematic parameters will suffer from cross-instrument averaging. Furthermore, as the NSIDC product does not use AMSR2, spatial resolution is inherently lower than KIMURA dataset. For these reasons, the NSIDC dataset is not considered further here. The KIMURA dataset is the only PM-derived dataset we consider here; it provides daily gridded sea ice motion vectors covering the complete Southern Ocean, and its 60 km horizontal resolution is suitable for exploring Antarctic sea ice kinematics in this study.

135
The objectives of this study are to a) evaluate the KIMURA ice motion product separately for the ASC, DES and combined datasets by comparison with coincident drifting buoy-derived velocities, and b) investigate impacts of the observational time-scale of the ice motion products on represention of Antarctic sea ice kinematic magnitude. This study will provide characterisation of the satellite sensor configuration and observational time-scale on the accuracy of derived sea ice motion, 140 as well as the ability to represent Antarctic sea ice kinematic parameters.

PM-derived sea ice motion
The KIMURA sea ice motion products (Kimura et al., 2013) used here were produced at 60×60 km resolution on a regular 145×145 grid covering the entire Southern Ocean, with daily data from 2012-06-23 to 2020-06-01. These were calculated by 145 applying the MCC method to 36 GHz, 10 km resolution AMSR2 TB images (both vertical and horizontal polarization, ASC and DES). Daily TB composites, i.e., the input imagery for the MCC algorithm, were obtained from the Arctic Data archive System (ADS) of the Japanese National Institute of Polar Research (NIPR). The Kimura et al. (2013) dataset combines both the daily ASC and DES sea ice motion datasets by taking their average. We also produce and analyze daily sea ice motion vectors computed for the shorter time-scale (24 h) ASC and DES swath composites.

In situ buoy-derived sea ice motion
Velocities derived from buoy positions are suitable for validating satellite-derived sea ice motion (Hoeber and Marianne, 1987;Heil and Allison, 1999). Here we evaluate the three KIMURA datasets using three buoys deployed on ice floes in the Weddell and Ross seas (Table 1 and Fig. 2). These have been named the southern Weddell Sea buoy (Fig. 2, red), the central Weddell Sea buoy (Fig. 2, green) and the Ross Sea buoy (Fig. 2, orange). In the Weddell Sea, several drifting sea ice buoys 155 were deployed in 2018 (Schröder, 2018). Only two of these are analyzed in this paper as the buoys were deployed in mesoscale arrays, hence at sub-grid scale to the KIMURA data. Others in this deployment were near the coast where the https://doi.org/10.5194/tc-2021-316 Preprint. Discussion started: 28 October 2021 c Author(s) 2021. CC BY 4.0 License.
were deployed as part of the Polynyas, Ice Production, and seasonal Evolution in the Ross Sea (PIPERS) cruise during austral autumn of 2017 (Tison et al., 2020). Four were excluded due to coastal proximity, and another one due to a short time series. Here we analyze only one of the three remaining buoys as they shared similar trajectories.  passes. This is particularly evident in the ASC case, where a strong discontinuity is observed at this latitude, owing to the oblique, forward-looking viewing geometry of the instrument. In both ASC and DES cases, more satellite swaths will be merged together at higher latitude (i.e., a higher degree of temporal "smearing" is expected at higher latitudes, with potential detriment to derived motion products). By extension, fewer swaths will be merged at lower latitudes, (i.e., at northern extremes, the DM product tends toward the S2S product, with reduced temporal smearing where u and v are daily sea ice velocities in the eastward and northward directions, respectively, while and represent eastward and northward components. We evaluate the magnitude of for the combined, ASC and DES datasets by

225
Our investigation into the source of the first problem revealed a second problem with the KIMURA dataset: an interval of exactly 24 h time separation between consecutive TB composites was assumed, but rarely correct. We analyze actual time separation between each pair of contiguous days. In order to characterise the variability of time separation, we calculate the standard deviation of time separation for both ASC and DES datasets. The result for September 2017 is displayed in Fig. 4 (all months similar, not shown here).
The mean time separation is close to 24 h throughout each month, however there is a region of high standard deviation around midnight UTC, which occurs in the Amundsen Sea for the ASC swaths (Fig. 4a) and the eastern Weddell Sea for the DES swaths (Fig. 4b). This indicates that the actual time separation of pixels in subsequent daily composites can regularly be up to 34 h, or as low as 14 h in these two areas. Thus, the resulting velocity is regularly incorrect by up to 40%, and the combined KIMURA dataset exhibits errors for these locations. However, any error is compensated by the following day's 235 velocity field, since the time intervals oscillate around 24 h. As such, it will not have a strong effect on long-term sea ice motion, but is highly detrimental when deriving sea ice kinematic parameters. To fix this problem, the actual time separation needs to be taken into account when producing daily sea ice motion velocity of KIMURA ASC, DES and combined datasets.
The appropriate correction has been performed for all remaining analyzes presented here. In summary, two problems have been discovered in the KIMURA datasets, and these two issues are rectified for the remainder of this work, as outlined above. Henceforth, the corrected KIMURA products are referred to as the KIMURAnew 245 datasets. The official release of the rectified KIMURA sea ice motion is a work in progress.

Validation of KIMURAnew datasets
The KIMURAnew dataset is compared to buoy-derived sea ice motion. The buoy start and end times needed to maximise correlation coincide exactly with the local (solar) times of corresponding satellite passes, i.e.,15:00, 0:00 and 8:00 UTC for ASC, DES and combined data, respectively, giving confidence in our methodology. Validation metrics are given in Table 2.    To investigate the hypothesis that DM-based daily sea ice motion vector performance is latitude-dependent, we calculate the three validation metrics in different latitude ranges along the buoy trajectories (Fig. 6). The Southern Weddell Sea buoy is excluded from this analysis due to insufficient latitudinal range. We find that the validation metrics for both remaining buoys 300 generally improve as they move further north (Fig. 6). For the central Weddell Sea buoy comparison, eastward velocity results ( Fig. 6 lower row solid lines) improve from south to north across the swath number discontinuity (74 o S), confirming our hypothesis that regions with fewer satellite swaths may yield improved performance for DM-based sea ice motion retrieval. However, northward ice velocity results in the central Weddell Sea case (dashed lines) exhibit a slight RMSD (Fig.   6f) deterioration from south to north. We attribute this unexpected deterioration to anisotropic TB heterogeneity, noting that 305 the heterogeneity in the northward component is lower than that in the eastward component (shown in

KIMURAnew-derived sea ice deformation
Accurate sea ice motion data is crucial for calculation of sea ice kinematic parameters. Here we investigate the effect of using the shorter time-scale ASC and DES sea ice motion products to retrieve sea ice , and compare it to those derived from the longer time-scale combined product.

Sea ice divergence
345 Fig. 8 illustrates per-pixel RMSD for the month of July 2017 derived from the KIMURAnew ASC, DES and combined datasets. The spatial mean circumpolar RMSD (Fig. 8) for the combined dataset (1.08×10 -6 s -1 ) is lower in magnitude than that of both the ASC (1.31×10 -6 s -1 ) and DES datasets (1.22×10 -6 s -1 ), indicating that the ASC and DES datasets can represent a higher magnitude of differential sea ice motion than the combined dataset. The magnitude of derived sea ice kinematics is demonstrated to be sensitive to time-scale, and we have shown that shorter time-scale input data produces higher sea ice 350 values.
To investigate the impacts of swath number on representing sea ice magnitude, we compare the July 2017 mean RMSD calculated by KIMURAnew ASC, DES and combined datasets in subregions north and south of the swath number discontinuity (74 o S for both ASC and DES, Fig. 2). The mean RMSD at higher latitudes is about 9%, 14% and 16% lower than at lower latitudes (i.e., north of 74 o S) for the ASC, DES and combined dataset, respectively. This indicates that DM-355 estimated sea ice RMSD are lower at higher latitudes, potentially due to a greater degree of smoothing due to more merged swath at higher latitudes, i.e., an undesirable bias may be introduced as a consequence of the observation configuration. While we have not investigated physical sea ice properties, which may also give rise to the lower DKP magnitude south of the swath discontinuity, we note that in the major Southern Ocean basins sea ice might be thinner and mechanically weaker at higher southern latitudes than that further north, since much of it has formed recently in coastal 360 polynyas (particularly in the south-western Ross and Weddell seas). As such, we would expect a greater magnitude of sea ice variability there. In this case, using an S2S dataset to derive sea ice motion may be required to remove observational bias.  . However, the central northern Weddell Sea stands out, exhibiting a region of high RMSD in the DES (Fig. 8b, red dashed circle), which is not mirrored in the ASC RMSD map. Such filamentary structures exist in the Weddell Sea during most wintertime months (not shown), and occur in ASC and DES maps equally, but the location is non-stationary. Given the non-stationary nature of this signal, we suggest that it is not an artifact of the satellite or analysis procedure. Noting the sun-370 synchronous observation platform, we suppose that this filamentary structure might be a response to external (oceanic or atmospheric) forcing with a ~24 h period, such as tidal currents or near-surface wind stress, but further research is beyond the scope of this paper. That this structure is a) not present in the ASC map, and b) muted in the combined map underscores the importance of considering observational time-scale and overpass time when interpreting maps of DKPs.

375
In this study, we firstly analyze DM-based Antarctic sea ice motion retrieval on sea ice kinematics by separating PM-derived KIMURAnew sea ice motion product into ASC and DES. Results show that performance of KIMURAnew datasets is a function of latitude, and that the time-scale of the composite dataset has crucial influence on retrieved sea ice kinematic magnitude. Here, we discuss the impacts of features of the satellite observational configuration on sea ice kinematics representation, and how to improve observations in the future.